[PVT] Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolution

paper: https://arxiv.org/abs/2102.12122
code: https://github.com/whai362/PVT/

1. Motivation

• 作者想实现一个纯净的，无卷积操作的transformer backbone，用于稠密任务。

As far as we know, exploring a clean and convolution-free Transformer backbone to address dense prediction tasks in computer vision is rarely studied.

• VIT的局限性，输入的特征图的分辨率较低，并且stride只能是16或者32，计算和内存开销都很大。

As far as we know, exploring a clean and convolution-free Transformer backbone to address dense prediction tasks in computer vision is rarely studied.

• (1) its output feature map has only a single scale with low resolution
• (2) its computations and memory cost are relatively high even for common input image size

图1为CNN网络，VIT网络以及作者提出的PVT，3种backbone的对比。

2. Contribution

• 本文提出了PVY金字塔视觉transformer，作为多种像素级别的任务的backbone，并且是不需要卷积的。

We propose Pyramid Vision Transformer (PVT), which is the first backbone designed for various pixel-level dense prediction tasks without convolutions. Combining PVT and DETR, we can build an end-to-end object detection system without convolutions and hand-crafted components such as dense anchors and non-maximum(NMS).

• PVT中实现了渐进收缩金字塔progressive shrinking pyramid以及 空间维度降维attention（SRA），提出的新结构可以用来减少资源的消耗，使得PVT对学习多尺度以及高分辨率的featmap更加的灵活，从而来实现dense predictions（典型工作就是检测和分割）。

We overcome many difficulties when porting Transformer to dense pixel-level predictions, by designing progressive shrinking pyramid and spatial-reduction attention (SRA), which are able to reduce the resource consumption of using Transformer, making PVT flexible to learn multi-scale and high-resolution feature maps.

• PVT可以应用于图像分类，目标检测，语义分割，实例分割等等。

We verify PVT by applying it to many different tasks, e.g., image classification, object detection, and semantic segmentation.

3. Method

3.1 Overall Architecture

3.2 Feature Pyramid for Transformer

​ 首先对于$H\times W\times 3$，首先根据patch size=4，分为16个patch，为了不改变图像带下，就将图像变为$\frac{H_{i-1}}{P_i}\times \frac{H_{i-1}}{P_i}\times (P_i^2{C}_{i-1})=\frac{H_{i-1}}{4}\times \frac{H_{i-1}}{4}\times (4\times 4\times 3)$，然后通过linear以及norm完成Patch Embedding操作，得到当前stage的$C_i$，最后得到的featmap 和PE结合，送入SRA中。

​ 但通过代码阅读后发现，linear实质上是直接通过卷积操作：input channel = 3，output channel = Ci，kernel大小为patch size的卷积核来实现的，相当于将原图$H\times W$通过卷积缩放为$\frac{H_{i-1}}{P_i}\times \frac{H_{i-1}}{P_i}$，然后通过卷积核的数量将通道变成$C_i$。从而可以得到$\frac{H_{i-1}}{P_i}\times \frac{H_{i-1}}{P_i}\times C_i$，对于stage1来说是$C_i=64$，从而在进入SRA中加入PE（绝对位置编码实现）。

3.3 Spatial-Reduction Attention

​ SRA相比较于普通的Scale-Product Attention，将K,V的空间维度进行了缩小，变成了$H_i{W}_i/R_i^2\times C_i$的大小，R为缩放因子（在5.2 SRA实现中用sr_ratio表示）这样子输出还是和Q的维度一致。

​ SRA的公式1,2与原multi-head attention基本类似，只是加入了$SR(x)$的reshape操作，如公式3所示。

3.3 Detailed settings of PVT series

​ 需要注意的是$C=[C_1,C_2,C_3,C_4]=[64,128,320,512]$以及$P=[P_1,P_2,P_3,P_4]=[4,2,2,2]$，每一个stage的featmap和原图之间的缩放倍数$[4,8,16,32]$，以及每一个stage中的encode含有多少个nhead，每一个stage中含有多少个encoder，从而分为了Tiny，Small，Medium，Large。

4. Experiments

4.1 Image Classification

4.2 Object Detection

4.3 Semantic Detection

4.4 Instance Segmentation

4.5 Pure Transformer Dense Prediction

4.5.1 PVT+ DETR


4.5.2 PVT+Trans2Seg

4.6 Ablation Experiments

4.6.1 Pre-trained Weights


4.6.2 Deeper vs. Wider

4.6.3 Computation Cost

5. Codes

5.1 PatchEmbed

class PatchEmbed(nn.Module):
    """ Image to Patch Embedding
    """
def __init__(self, img_size=224, =4, in_chans=3, embed_dim=64):
    super().__init__()
    img_size = to_2tuple(img_size)
    patch_size = to_2tuple(patch_size)

    self.img_size = img_size
    self.patch_size = patch_size
    assert img_size[0] % patch_size[0] == 0 and img_size[1] % patch_size[1] == 0, \
        f"img_size {img_size} should be divided by patch_size {patch_size}."
    self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
    self.num_patches = self.H * self.W
    self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
    self.norm = nn.LayerNorm(embed_dim)

def forward(self, x):
    B, C, H, W = x.shape

    x = self.proj(x).flatten(2).transpose(1, 2)
    x = self.norm(x)
    H, W = H // self.patch_size[0], W // self.patch_size[1]

    return x, (H, W)

5.2 Attention (SRA)

    class Attention(nn.Module):
        def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
            super().__init__()
            assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
    	self.dim = dim
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5
    
        self.q = nn.Linear(dim, dim, bias=qkv_bias)
        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)
    
        self.sr_ratio = sr_ratio
        if sr_ratio > 1:
            self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
            self.norm = nn.LayerNorm(dim)
    
    def forward(self, x, H, W):
        B, N, C = x.shape
        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
    	# sra
        if self.sr_ratio > 1:
            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
            x_ = self.norm(x_)
            kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        else:
            kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        k, v = kv[0], kv[1]
    
        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)
    
        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
    
        return x

5.3 model

@register_model
def pvt_tiny(pretrained=False, **kwargs):
    model = PyramidVisionTransformer(
        patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4], qkv_bias=True,
        norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
        **kwargs)
    model.default_cfg = _cfg()

    return model